Sandbox0
Self hosted sandbox runtime for running long lived AI agents with gVisor isolation, durable filesystems, incremental checkpoints and replaceable compute
Overview
Sandbox0 is an open source runtime for running AI agents inside isolated environments while preserving their working files between compute sessions. An agent can install dependencies, clone repositories and generate output, then pause and resume later with the same writable filesystem attached to fresh compute.
Its durable RootFS stores encrypted filesystem blocks in S3 compatible object storage. Incremental checkpoints publish only changed blocks, while resumed sandboxes fetch committed data as it is needed. The normal pause process preserves files but starts fresh processes when the sandbox returns. Experimental memory checkpoints can also retain supported process state, though they are not ready for routine production use.
Sandbox0 uses Nomad for scheduling and gVisor for workload isolation. Teams can manage sandboxes through Python, JavaScript, TypeScript and Go SDKs or the s0 command line interface. It also includes snapshots, forks, network policies, scoped credentials, SSH access, logs, metrics and API controls.
Self hosting is an infrastructure project, not a simple container installation. A production deployment requires dedicated Linux Nomad clients, PostgreSQL, S3 compatible storage, private DNS, mutual TLS and specific kernel and networking capabilities. Those workers cannot also run unrelated Nomad jobs. The project is moving quickly and currently lacks a published security policy, which matters for software that executes untrusted code. It best suits experienced platform teams building persistent coding agent infrastructure.
Use Cases
Specific ways to use Sandbox0 for your workflow.
Deployment Strategy
Recommended ways to host Sandbox0 in your own environment.